Intriguing Properties of Generative Classifiers

Dear reviewer QCiG,

We’d like to thank you for your time and review. We’re happy that you appreciated the historical context of generative vs. discriminative classifiers and found our approach a “useful contribution” with “excellent soundness” and “interesting findings & discussion”. We address the questions and concerns you raised below.

1. “The general direction of the findings is not terribly surprising, although the exact degree of the results and the precision of this study is still useful I think”:

We are happy to hear that the reviewer appreciated the “precision of the study” we undertook in this paper. We believe (as also mentioned by reviewer e2eW) that “the significance of the empirical results are the key novelty of the paper”. In particular, we’d like to point out a few aspects where our empirical results are a significant and novel contribution, which can be assessed independently of prior expectations:

• This is the first study to compare modern generative and discriminative classifiers with large-scale human perceptual data on 17 challenging datasets
• The method sets a new SOTA in shape bias (first-ever human-like shape bias by a ML model)
• We demonstrate that generative classifiers achieve SOTA in human error alignment, even though deep learning models of human visual perception (and deep learning models in general) are currently dominated by discriminative classifiers
• We expect a number of different follow-up directions will emerge from this work: in the ML community, harnessing generative pre-training in combination with faster inference (perhaps even via distillation as you suggested?) for various downstream tasks; in the cognitive science community, studying perceptual illusions with generative models; in the neuroscience community, building more accurate models of how generative and discriminative processes might be combined in biological systems like the primate brain.

2. Distillation:

That is a very interesting suggestion! Yes, we think it might be possible to distill the scores (i.e., variational lower bound estimates) from Imagen into a discriminative classifier. Next to the computational cost of generating thousands of samples, another practical challenge might be the architectural mismatch between the diffusion model U-Net and most discriminative models; for example a diffusion model has cross attention between the input text and image while a two tower model like CLIP does not. So we are not sure how well it would work in practice, but think it would be an interesting direction to explore in future work. More broadly, our experiment with adding diffusion-style noise to a supervised classifier (Section 4) shows that lessons from diffusion models can be used to improve the shape bias of a discriminative model; which is in line with https://arxiv.org/abs/2304.08466 using a combination of synthetic and natural images to improve ImageNet accuracies.

3. Probing such that discriminative classifiers have advantage over generative classifiers:

We note that our experimental evaluation is unbiased since we are using a very broad existing benchmark of 17 datasets that is, in fact, a benchmark designed for discriminative models and not generative models. The results presented in the manuscript are, therefore, in no way cherry-picked in favour of generative models.

Within our experiments, we already see some places where discriminative models are better than generative counterparts: for example, as we observed in Section 3.2 and discuss in Section 5, generative models have a hard time classifying or generating rotated images while discriminative models are easily able to do this. Similarly, in the shape-vs-texture bias experiment, if the objective is to recognize the correct texture, discriminative models have a greater advantage over discriminative models. Finally, as mentioned in Jaini & Clark, 2023, on OCR like datasets (e.g. MNIST, and SVHN), discriminative models are able to achieve near-perfect accuracy while generative models still struggle (Stable Diffusion gets ~18% and Imagen gets ~79% on MNIST).

Thanks again for your questions and thoughts!

Dear reviewer e2eW,

Thanks for your time and thorough review, we’re happy to hear that you appreciated the “extensive comparison”, the paper’s “excellent” soundness and presentation, and the “significance of the empirical results”. We address the questions and concerns you raised below.

More discussion regarding ViT-22B-384:

That’s a good point, ViT-22B-384 is indeed a very competitive classifier (perhaps not surprising given that it is a large vision transformer with 22B parameters trained on 4B images for object recognition). In response to your comment, we’ve now added some discussion and made sure to reference the model in each of the empirical results sections (see 3.1, 3.2 and 3.3). For instance, in Section 3.2, we now write “We find that Imagen and Stable Diffusion achieve an overall accuracy that is close to human-level robustness (cf. Figure 3) despite being zero-shot models; these generative models are outperformed by some very competitive discriminative models like ViT-22B achieving super-human accuracies.”; Section 3.1 states “we report that all three generative classifiers significantly outperform ViT-22B (Dehghani 2023), the previous state-of-the-art method in terms of shape bias, even though all three models are smaller in size, trained on less data, and unlike ViT-22B were not designed for classification.” and Section 3.3 states “While a substantial gap towards human-to-human error consistency remains, Imagen shows the most human-aligned error patterns, surpassing previous state-of-the-art (SOTA) set by ViT-22B, a large vision transformer (Dehghani 2023).”

Different resizing resolutions:

Imagen uses lower input resolution than Stable Diffusion because Imagen is a cascaded diffusion model and we only use the first stage image generator, not the superresolution models. The low resolution certainly is a disadvantage for Imagen in terms of performance, which we think makes its strong results all the more impressive. We have added the following note in the corresponding methods section on preprocessing (Section 2) to explain this choice: “ We preprocess the 17 datasets in the model-vs-human toolbox by resizing the images to $64 \times 64$ resolution for Imagen, $256 \times 256$ for Parti, and $512 \times 512$ for SD since these are the resolutions for the each of the base models respectively.”.

Design choices:

We used exactly the same design choices as mentioned in Clark & Jaini (2023) for our experiments. We have now added a discussion on this in Appendix A.

L2 loss function:

We use the L2 loss function for diffusion-based models since it approximates the diffusion variational lower bound (see Eq. 2) and thus results in a Bayesian classifier. Furthermore, both Stable Diffusion and Imagen are trained with the L2 loss objective. Thus, a different loss function will no longer result in a Bayesian classifier and will not work well due to differences from the training paradigm. We have added a discussion on the motivation for this choice in Appendix A.

Thanks again for your questions and time!

Dear reviewer tjak,

Thank you for your review and comments / suggestions! We’re happy to hear you found the presentation and contribution “excellent” and the approach “well-executed” and “both important and timely”.

1. Rich semantic representation:

We agree that attributing findings to the models' generative nature or their rich semantic representations is interesting! We first note that CLIP is a contrastive based method that was trained to match images with captions; thus we don’t consider CLIP to be a discriminative model baseline; it was not trained for specific downstream tasks like object classification, but instead performs zero-shot classification. CLIP performs worse than Stable Diffusion in OOD accuracy and similarly in error consistency when CLIP uses a single text prompt rather than an ensemble of 80 different prompts. To make this clearer, we have updated Table 1 to include 1-prompt CLIP results which are directly comparable with the 1-prompt results for all generative classifiers in the paper.

By semantic representation, we assume you’re referring to the text representations? If so, this is unlikely to explain the high performance of the generative classifiers. While using a powerful text model may be important in settings involving complex language descriptions, the input prompts we use are very simple. It doesn’t seem likely to us that a 10b-parameter text encoder like T5-XXL offers much of an advantage when the prompt is “a bad photo of a dog.” In other words, the challenge of the out-of-distribution datasets we use is in the visual understanding needed to solve the tasks, not in language understanding. That said, we agree that a perfect comparison would include the same language encoder for every model, thus we have added the following to the paper’s limitations section: “Furthermore, different models use different language encoders which may be a confounding factor”.

2. clustering analysis:

Indeed, our Figure 6 showed CLIP in the ‘model cluster’ rather than the ‘human cluster’ because we sorted the models manually and admittedly ad-hoc by model type, starting with CNNs etc. In response to your concern, we have updated the plot with a more principled sorting method: we now sort by average error consistency with human observers - thus the resulting clusters are now based on actual data, rather than manual sorting. There are still broadly two clusters that emerge - one big cluster encompassing most models, and the other encompassing humans, generative classifiers, CLIP and a few other models. Consequently, we have also updated the annotation from “Generative classifiers are in the human cluster, not the model cluster” to “Generative classifiers show high consistency with humans and lower consistency with ‘traditional’ models” (since it’s no longer just a ‘human cluster’). We believe the updated plot (Figure 6) better reflects the empirical finding, thanks for your comment!

3. Open sourcing code:

We agree with the reviewer and are working on open-sourcing code as soon as possible. We would additionally like to note that most components of our paper can be replicated already using existing open-source repositories by various different authors and repos:

The model-vs-human toolbox that contains all datasets, benchmarks, and various models: https://github.com/bethgelab/model-vs-human

Zero-shot classification code for Stable Diffusion: https://github.com/diffusion-classifier/diffusion-classifier

Finally, we have provided a high level of detail for reproducibility. A core experiment from our study (increased shape bias from diffusion-style training) has already been successfully replicated by an independent third party researcher based on our methods section. We provide the link to this independent replication with the caveat that they link to our publicly available non-anonymous version of the paper that might compromise the double-blind review process. Please do not visit the link if that is a concern: https://github.com/paulgavrikov/diffusion-noise-cls-pytorch

4. Only considering final, resulting image:

Maybe there is a misunderstanding: in our experiments we always evaluate the score based on the final denoised image (for example please see Figure 3 in Clark & Jaini 2023 on which our approach is based). We also do not perform sampling but just run the forward diffusion process to get scores. The averaging in Eq.2 is over the noise level - we sample multiple noise levels to noise the target image and then denoise it to get the relevant scores. We have added an algorithm box in Appendix A to make this clearer.

5. Citing Golan, Raju & Kriegeskorte (2020):

We have added a citation to the controversial stimuli paper in Section 3 (“These findings appear consistent with the MNIST results by Golan et al. (2020) reporting that a generative model captures human responses better than discriminative models.”)

Thanks again for your detailed review!